This specification relates to parallelizing the training of convolutional neural networks.
Convolutional neural networks generally include two kinds of neural network layers, convolutional neural network layers and fully-connected neural network layers. Convolutional neural network layers have sparse connectivity, with each node in a convolutional layer receiving input from only a subset of the nodes in the next lowest neural network layer. Some convolutional neural network layers have nodes that share weights with other nodes in the layer. Nodes in fully-connected layers, however, receive input from each node in the next lowest neural network layer.